VLM-Mapper: Multimodal vision-language models for prior guided lane-level HD map construction

Abstract

The automated construction of High-Definition (HD) maps from remote sensing data is essential for modern intelligent transportation systems and spatial data infrastructure. While imagery provides a scalable solution for lane-level HD map construction, traditional discriminative models often struggle in complex scenarios because of their reliance on local visual features and limited use of external geographic context. To address these limitations, this paper introduces VLM-Mapper, a framework that reframes lane-level HD map construction as a multimodal sequence-generation task based on visual–language alignment and prior-conditioned token generation. VLM-Mapper systematically integrates high-resolution aerial imagery with tokenized geographic prior prompts, including standard map metadata and boundary entry points, to guide HD map construction. Furthermore, to mitigate the spatial inconsistencies and count errors of autoregressive decoding, we implement a two-stage post-training pipeline. This strategy establishes spatial alignment via Supervised Fine-Tuning (SFT), followed by a Reinforcement Learning (RL) phase using Group Relative Policy Optimization (GRPO) with a custom geometry-aware and numerical reward function. Experimental results on the large-scale OpenSatMap dataset show that VLM-Mapper achieves state-of-the-art performance and improves the robustness and topological consistency of HD map construction across large-scale aerial imagery.